Huawei is gearing up to launch its latest AI accelerator, the Ascend 910C, which has just entered production and is generating quite a buzz in the industry.
In what can only be described as a technological revolution, China’s AI sector is booming. Companies like DeepSeek have introduced models such as the R1, escalating the demand for robust computing solutions. Among the Chinese AI powerhouses, Huawei is particularly noteworthy. Their soon-to-be-released Ascend 910C promises to stir up competition, particularly as it poses a formidable challenge to NVIDIA’s H100 in the Chinese market. According to insights from @ohlennart, the details of the Ascend 910C suggest it will be a serious contender.
Huawei’s strategy with the Ascend 910C, outlined in the dissemination by @ohlennart, shows a distinct approach compared to NVIDIA. The design relies on two separate silicon interposers connected through an organic substrate—opting for a more conventional method. Essentially, this combines two Ascend 910B chips to amplify performance. The chip is rumored to achieve 800 TFLOP/s at FP16 and boast a memory bandwidth of up to 3.2 TB/s, which is nearly comparable to the performance of the H100.
Underneath the hood, Huawei is reportedly employing the 7nm technology from TSMC and SMIC. Before the imposition of export restrictions, Huawei placed significant 7nm orders with TSMC, hinting that much of their production will stem from this collaboration. Meanwhile, SMIC, China’s native manufacturer, is expected to contribute significantly as well, with a notable monthly wafer output of around 50,000. This dual sourcing strategy positions Huawei to achieve its ambitious goal of shipping millions of AI chips.
The Ascend 910C represents a milestone in China’s domestic AI capabilities and a promising leap forward. Yet, globally, Huawei still trails in computing power by a factor of 10 to 20 times. Nonetheless, firms like DeepSeek are indicative of China’s ability to innovate and find unique solutions that mitigate these discrepancies. With such talent and strategic maneuvering, the global AI landscape is expected to become significantly more competitive in the near future.
